Pages: [1]   Go Down

Author Topic: Compare image quality (from statistical point of view)  (Read 1281 times)

divo

  • Newbie
  • *
  • Offline Offline
  • Posts: 1
Compare image quality (from statistical point of view)
« on: August 03, 2011, 02:14:40 am »

First of all I want to say hello and to thank to all users so that this forum is so helpful :)

I need your help to determine statistical method of image quality assessment. What I got is few tone-mapped HDR photos (same scene, different tone-mapping operators) and I want to determine which picture contains highest useful information in shadows, highlights and midtones simultaneously. I think dynamic range and contrast will be the best (and simpliest?) variables but I'm open to your suggestions. I don't want to measure color accuracy, distortion, sharpness, noise etc. because my assumption is this are constant or insignificant data.

You might ask why I need those things. It's a college project so... well... I don't know ;] but if longer to think about it this might be useful for example in CCTV industry, when small file size is important and have to be stored in traditional, 8bpc, low dynamic range format but we want to store as many information as possible for future text or face recognition. Of course we can look at image, maybe process it a little and we can say that one is better than other, but this is not a case from "scientific" pov. End of digression :)

My method:
I've made an attempt to determine three 32x32 px samples with some shadows, highlights and midtones from each image (same area for each image). Next convert RGB values of each pixel to luminosity and put this values to Excel for further calculation. Now I need your help to determine what statistical tools should I use to find out which sample (and as a consequence whole image) contain max data and how to interpret figures.

I hope I've comprehensible illustrated the problem. If not, and you think you can help, please ask. Any directions and comments will be very useful for me.

Best Regards,
Pawel

Logged

Bart_van_der_Wolf

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 8913
Re: Compare image quality (from statistical point of view)
« Reply #1 on: August 03, 2011, 06:42:09 am »

First of all I want to say hello and to thank to all users so that this forum is so helpful :)

I need your help to determine statistical method of image quality assessment. What I got is few tone-mapped HDR photos (same scene, different tone-mapping operators) and I want to determine which picture contains highest useful information in shadows, highlights and midtones simultaneously. I think dynamic range and contrast will be the best (and simpliest?) variables but I'm open to your suggestions. I don't want to measure color accuracy, distortion, sharpness, noise etc. because my assumption is this are constant or insignificant data.

Hi Pawel,

You can prove anything with statistics, so you'll need to be clear on what constitutes 'quality' (and make sure you can explain why). Then find a statistical method that gives a good discrimination between a 'optimal' and 'poor' Figure of Merit (FOM).

Quote
You might ask why I need those things. It's a college project so... well... I don't know ;] but if longer to think about it this might be useful for example in CCTV industry, when small file size is important and have to be stored in traditional, 8bpc, low dynamic range format but we want to store as many information as possible for future text or face recognition. Of course we can look at image, maybe process it a little and we can say that one is better than other, but this is not a case from "scientific" pov. End of digression :)

Well, we're not supposed to do your homework, so I can only give some suggestions (assuming part of the project is the process of figuring out stuff). Get hold of a copy of  book by Gonzalez & Woods, "Digital Image Processing". Check out chapter 11.3.3 about quantifying texture content.

Quote
My method:
I've made an attempt to determine three 32x32 px samples with some shadows, highlights and midtones from each image (same area for each image). Next convert RGB values of each pixel to luminosity and put this values to Excel for further calculation. Now I need your help to determine what statistical tools should I use to find out which sample (and as a consequence whole image) contain max data and how to interpret figures.

Maybe, just maybe, homogeneity is a parameter that could be useful. I'm not saying it is, because it depends on the subject being imaged. One thing you won't have to worry about is noise, assuming a proper HDR image with high S/N ratio at all luminosity levels.

Do let us know what progress is made, because it might help others looking for similar solutions.

Cheers,
Bart
Logged
== If you do what you did, you'll get what you got. ==
Pages: [1]   Go Up